I did a simple test for my MAX (Machine Assisted VFX) Workflow.

I made an 8K UV texture (because why not) and mapped it onto a 3D Sphere in Nuke.

The Sphere rotates on its own Y-axis over 96 frames.

Then I duplicated the Sphere, gave it a new grey checkerboard texture, and animated it floating in front of the UV-textured one, like a confused NPC from Oblivion getting in the way.

The goal?

Use CopyCat inference to remove the grey Sphere.

Since I had a clean render of the textured Sphere, I fed that in as Ground Truth.

Trained CopyCat on every 12th frame using the default values.

Here's what I got.

0:00
/0:18